Anthropic's Fable 5 Unbanned, Sonnet 5 Arrives Amidst Performance and Cost Scrutiny

Anthropic’s Fable 5 model, previously restricted by the Secretary of Commerce on June 12th due to concerns regarding its ability to identify security vulnerabilities, has been unbanned. The lifting of export controls on June 26th follows Anthropic’s close coordination with the US government, where the company committed to proactive security risk detection, protocol development, and reporting of malicious activity. This pivotal decision now allows for the export, re-export, and in-country transfer of Fable 5, enabling its integration into API-driven services. This development coincides with the release of Sonnet 5, a new model that Anthropic positions as its most agentic Sonnet to date, boasting enhanced capabilities in planning, tool utilization, and autonomous operation previously associated with more expensive, larger models.

Despite Anthropic’s claims that Sonnet 5 offers performance comparable to Opus 4.8 at lower prices and significant improvements over its predecessor, Sonnet 4.6, initial analyses and real-world testing present a nuanced perspective. Benchmarks from Cursor Bench indicate Sonnet 5 generally underperforms and is more expensive than GPT-5.5 equivalents. Artificial Analysis’s cost metrics highlight Sonnet 5 as the most expensive model run through their benchmark at $6,000, consuming substantially more tokens than Opus or GPT-5.5, leading to slower task completion and increased real-world costs despite its lower per-token introductory pricing. While Sonnet 5 demonstrates advanced agentic behaviors, including sub-agent orchestration and task decomposition—qualities previously observed in Fable 5—its execution is often hampered by a perceived lack of sufficient intelligence, resulting in prolonged runtimes, elevated expenses, and a tendency to loop inefficiently. Reports further detail a bug in cloud.ai accounts where Sonnet 5 leaks extensive internal ‘thinking traces,’ exacerbating token waste. The model’s safety profile also shows a trade-off, with a higher refusal rate for malicious requests but a diminished success rate for benign-yet-suspicious queries compared to Sonnet 4.6. Consequently, while Sonnet 5’s orchestration capabilities suggest a niche role as a tool for more intelligent, overarching orchestrators like Fable 5 or future GPT models, the prevailing sentiment among developers leans towards continuing with GPT-5.5 and Opus for daily workloads, with keen anticipation for Fable 5’s full re-entry into the market.